WIP: handle zero_tangent from cyclic data structures v2 Via premapping

[oxinabox]

This is an alternative to #654
I think it is more promising and should minimize the type instability outside of cyclic locations

Basically here is the idea:

1. Identify all objects that have multiple references to them
2. Go through and construct tangents for everything;
3. save the tangents for things with >1 reference into an IdDict keyed by primal
4. when constructing tangent fields for things that have >1 reference to them, initially treat them as per undef elements, so set the tangent value as also undef or to a ZeroTangent
5. go back through the tangent object and set the values for things with multiple references

This will also handle aliases outside of loops correctly.

[oxinabox]

no this plan won't work right as as.
You can't always mute in the references after the fact.
Because a immutable type might have fields that are references.

I think we are going to need to topo sort this or something.
I am going to draw some pictures

[oxinabox]

ok, so cycles are only possible with mutation.
So the kinda extreme version of what we can do is

1. do cycle detection + reference counting
2. break the graph at (at least one) mutable struct field/array element; creating a Tangent putting in a ZeroTangent for that value, but otherwise constructing as per step 3 below; and adding to a shared primal indexed Dict of of lists of pending work a function that will mutate the tangent to attach a tangent computed later.
3. Construct tangent recursively, For any field with multiple references and that is not isbits (these two checks are an optimization and could be skipped) check in a primal keyed dict of already constructed tangents and if it is found use that, otherwise add it to that.
4. run all the pending work from step 2 to relink up the graph.